---

license: apache-2.0

base_model: Ilides/Spartan-1.2-2B-MLX-8bit

tags:

- gguf

- qwen3.5

- spartan

- ilides

- llama-cpp

language:

- en

- es

pipeline_tag: text-generation

---

Spartan 1.2 2B (GGUF F16)

GGUF export of Spartan 1.2 MLX for llama.cpp, LM Studio, Ollama and Jarvis on any platform (not only Apple Silicon).

> For the latest MLX build on Mac, see Spartan 1.3.

---

Files

| File | Size | Description |

|------|------|-------------|

| Spartan-1.2-2B-F16.gguf | 3.5 GB | Float16 weights (320 tensors) |

| chat_template.jinja | 8 KB | Patched Qwen3.5 template (no tool-loop crash) |

| README.md | — | This card |

---

Model specs

| Property | Value |

|----------|-------|

| Source | Spartan 1.2 MLX (bf16 fused, pre-quantize) |

| Architecture | Qwen3_5ForConditionalGeneration |

| Parameters | 1,881,824,512 (~1.88B) |

| Layers | 24 |

| Hidden size | 2048 |

| FFN size | 6144 |

| Attention heads | 8 (KV heads: 2) |

| Vocab / embedding | 248,320 / 2048 |

| Max context | 262,144 (model native) |

| GGUF quantization | F16 (no Q4_K_M in this release) |

| Converter | llama.cpp convert_hf_to_gguf.py --no-mtp |

Behavior: identical to Spartan 1.2 MLX — anti-loop patch, agentic tool format, workspace-relative path training. Functional tests were run on the MLX source; see that model card for results.

---

Export details (real log)

Source:  output/Spartan-1.2-2B-MLX (bf16 fused)
Output:  Spartan-1.2-2B-F16.gguf
Tensors: 320
Size:    3.76 GB on disk → 3.5 GB in repo
Tool:    vendor/llama.cpp convert_hf_to_gguf.py --outtype f16 --no-mtp

Q4_K_M quantization was not generated (llama-quantize not built at export time).

---

Functional tests

Validated on the MLX source model (2026-08-03). Expected same behavior in GGUF:

| Test | Result |

|------|--------|

| Chat (async/await) | ✅ |

| Coding (Rust binary search) | ❌* |

| Agentic grep TODO | ✅ |

| Agentic Write relative path | ❌** |

| Workspace path safety | ✅ |

\* See Spartan 1.2 MLX card for details.

\** Model may output code prose instead of [Tool: Write] on creation prompts.

Re-run tests on MLX:

python scripts/validate_spartan.py --model ../cortex/models/Spartan-1.2-2B-MLX-8bit

---

Load

llama.cpp

llama-cli -m Spartan-1.2-2B-F16.gguf -p "Hello" -n 256

LM Studio

1. Download Spartan-1.2-2B-F16.gguf
2. Load in LM Studio (GGUF backend)
3. Set system prompt (see below)
4. Context: 8192 recommended
5. Thinking: OFF

Ollama

# Create Modelfile pointing to the GGUF, or import via LM Studio

---

System prompt

Use the Spartan agentic prompt — workspace-relative paths, anti-loop:

You are Spartan 1.2, an agentic coding assistant created by ilides.

Tools: Bash, Grep, Read, Write, StrReplace, Glob.
Format: [Tool: ToolName] {"arg": "value"}

WORKSPACE ONLY:
- All files go in the open workspace. Relative paths only (index.html).
- Never /Users/..., /Desktop/..., /Applications/..., or any absolute path.
- Never request editing outside the workspace. No cd to parent or Desktop.

ANTI-LOOP: never repeat cd/bash; use Write for files; one simple stack per task.

Full prompt from repo: python scripts/print_system_prompt.py agentic_v11

---

Intended use

• Jarvis (spartan-1.2-2b-mlx / GGUF loader)
• LM Studio on Windows/Linux/Mac (CPU/GPU via llama.cpp)
• Ollama local inference
• Cross-platform deployment where MLX is not available

---

Known limitations

• F16 only: 3.5 GB — no quantized Q4 variant in this repo
• Same 2B limits as MLX 1.2: weak on long agent chains, HTML quality, reliable tool syntax
• Thinking tags: disable in host
• Chat template patch required: included as chat_template.jinja — fixes "No user query found" crash on tool-only message tails
• 1.3 not exported: workspace patch exists only in MLX 1.3

---

Lineage

Qwen3.5-2B → Spartan 2B → Spartan 1.1 → Spartan 1.2 MLX → Spartan 1.2 GGUF (this file)

---

License

Apache 2.0.